Reinforcement learning (RL) for LLM reasoning has rapidly emerged as a prominent research area, marked by a significant surge in related studies on both algorithmic innovations and practical applications. Despite this progress, several critical challenges remain, including the absence of standardized guidelines for applying RL techniques and a fragmented understanding of their underlying mechanisms. In addition, inconsistent experimental settings, variations in training data, and differences in model initialization have led to conflicting conclusions, obscuring the key characteristics of these techniques and creating confusion among practitioners when selecting appropriate techniques. This paper systematically reviews widely adopted RL techniques through rigorous reproductions and isolated evaluations within a unified open-source framework. We analyze the internal mechanisms, applicable scenarios, and core principles of each technique through fine-grained experiments, including datasets of varying difficulty, model sizes, and architectures. Based on these insights, we present clear guidelines for selecting RL techniques tailored to specific setups and provide a reliable roadmap for practitioners navigating the RL for the LLM domain. Finally, we show that a minimalist combination of two techniques can unlock the learning capability of critic-free policies with a vanilla PPO loss. The results demonstrate that our simple combination consistently improves performance, surpassing strategies such as GRPO and DAPO.

Transformer decoders have achieved strong results across tasks, but the memory required for the KV cache becomes prohibitive at long sequence lengths. Although Cross-layer KV Cache sharing (e.g., YOCO, CLA) offers a path to mitigate KV Cache bottleneck, it typically underperforms within-layer methods like GQA. To understand the root cause, we investigate the information flow of keys and values of the top-layers. Our preliminary reveals a clear distribution: values are predominantly derived from the bottom layer, while keys draw more information from both bottom and middle layers. Building upon this, we propose FusedKV, whose top-layer KV caches are a learnable fusion of the most informative ones from the bottom and middle layers. This fusion operates directly on post-RoPE keys, preserving relative positional information without the computational cost of re-applying rotary embeddings. To further improve efficiency, we propose FusedKV-Lite, an cross-layer sharing approach, where top-layer KV caches are directly derived from the bottom-layer values and the middle-layer keys. Compared to FusedKV, FusedKV-Lite reduces I/O overhead at the cost of a slight increase in perplexity. In experiments on LLMs ranging from 332M to 4B parameters, our proposed method reduce 50\% cache memory while achieving lower validation perplexity than the standard Transformer decoder, establishing it as a memory-efficient, high-performance architectural alternative. We have made our Triton implementation available.

We introduce ROLL, an efficient, scalable, and user-friendly library designed for Reinforcement Learning Optimization for Large-scale Learning. ROLL caters to three primary user groups: tech pioneers aiming for cost-effective, fault-tolerant large-scale training, developers requiring flexible control over training workflows, and researchers seeking agile experimentation. ROLL is built upon several key modules to serve these user groups effectively. First, a single-controller architecture combined with an abstraction of the parallel worker simplifies the development of the training pipeline. Second, the parallel strategy and data transfer modules enable efficient and scalable training. Third, the rollout scheduler offers fine-grained management of each sample's lifecycle during the rollout stage. Fourth, the environment worker and reward worker support rapid and flexible experimentation with agentic RL algorithms and reward designs. Finally, AutoDeviceMapping allows users to assign resources to different models flexibly across various stages.

Recent advances in Vision-Language-Action models (VLAs) have expanded the capabilities of embodied intelligence. However, significant challenges remain in real-time decision-making in complex 3D environments, which demand second-level responses, high-resolution perception, and tactical reasoning under dynamic conditions. To advance the field, we introduce CombatVLA, an efficient VLA model optimized for combat tasks in 3D action role-playing games(ARPGs). Specifically, our CombatVLA is a 3B model trained on video-action pairs collected by an action tracker, where the data is formatted as action-of-thought (AoT) sequences. Thereafter, CombatVLA seamlessly integrates into an action execution framework, allowing efficient inference through our truncated AoT strategy. Experimental results demonstrate that CombatVLA not only outperforms all existing models on the combat understanding benchmark but also achieves a 50-fold acceleration in game combat. Moreover, it has a higher task success rate than human players. We will open-source all resources, including the action tracker, dataset, benchmark, model weights, training code, and the implementation of the framework at https://combatvla.github.io/.

The emergence of Large Language Models (LLMs) has necessitated the adoption of distributed training techniques, involving the deployment of thousands of GPUs to train a single model. Unfortunately, the efficiency of large-scale distributed training systems is often suboptimal due to the increased likelihood of hardware errors in high-end GPU products and the heightened risk of network traffic collisions. Specifically, GPUs involved in the same job require periodic synchronization to exchange necessary data, such as gradients, parameters, or activations. As a result, any local hardware failure can disrupt training tasks, and the inability to swiftly identify faulty components leads to a significant waste of GPU resources. Moreover, prolonged communication due to traffic collisions can substantially increase GPU waiting times. To address these challenges, we propose a communicationdriven solution, namely the C4. The key insights of C4 are twofold. First, the load in distributed training exhibits homogeneous characteristics and is divided into iterations through periodic synchronization, therefore hardware anomalies would incur certain syndrome in collective communication. By leveraging this feature, C4 can rapidly identify the faulty components, swiftly isolate the anomaly, and restart the task, thereby avoiding resource wastage caused by delays in anomaly detection. Second, the predictable communication model of collective communication, involving a limited number of long-lived flows, allows C4 to efficiently execute traffic planning, substantially reducing bandwidth competition among these flows. The C4 has been extensively deployed across real-world production systems in a hyperscale cloud provider, yielding a significant improvement in system efficiency, from 30% to 45%. This enhancement is attributed to a 30% reduction in error-induced overhead and a 15% reduction in communication costs.

Reinforcement learning for LLM reasoning has rapidly emerged as a prominent research area, marked by a significant surge in related studies on both algorithmic innovations and practical applications. Despite this progress, several critical challenges remain, including the absence of standardized guidelines for employing RL techniques and a fragmented understanding of their underlying mechanisms. Additionally, inconsistent experimental settings, variations in training data, and differences in model initialization have led to conflicting conclusions, obscuring the key characteristics of these techniques and creating confusion among practitioners when selecting appropriate techniques. This paper systematically reviews widely adopted RL techniques through rigorous reproductions and isolated evaluations within a unified open-source framework. We analyze the internal mechanisms, applicable scenarios, and core principles of each technique through fine-grained experiments, including datasets of varying difficulty, model sizes, and architectures. Based on these insights, we present clear guidelines for selecting RL techniques tailored to specific setups, and provide a reliable roadmap for practitioners navigating the RL for the LLM domain. Finally, we reveal that a minimalist combination of two techniques can unlock the learning capability of critic-free policies using vanilla PPO loss. The results demonstrate that our simple combination consistently improves performance, surpassing strategies like GRPO and DAPO.

Continual Pre-Training (CPT) on Large Language Models (LLMs) has been widely used to expand the model’s fundamental understanding of specific downstream domains (e.g., math and code). For the CPT on domain-specific LLMs, one important question is how to choose the optimal mixture ratio between the general-corpus (e.g., Dolma, Slim-pajama) and the downstream domain-corpus. Existing methods usually adopt laborious human efforts by grid-searching on a set of mixture ratios, which require high GPU training consumption costs. Besides, we cannot guarantee the selected ratio is optimal for the specific domain. To address the limitations of existing methods, inspired by the Scaling Law for performance prediction, we propose to investigate the Scaling Law of the Domain-specific Continual Pre-Training (D-CPT Law) to decide the optimal mixture ratio with acceptable training costs for LLMs of different sizes. Specifically, by fitting the D-CPT Law, we can easily predict the general and downstream performance of arbitrary mixture ratios, model sizes, and dataset sizes using small-scale training costs on limited experiments. Moreover, we also extend our standard D-CPT Law on cross-domain settings and propose the Cross-Domain D-CPT Law to predict the D-CPT law of target domains, where very small training costs (about 1\% of the normal training costs) are needed for the target domains. Comprehensive experimental results on six downstream domains demonstrate the effectiveness and generalizability of our proposed D-CPT Law and Cross-Domain D-CPT Law.

Fail-slows, or stragglers, are common but largely unheeded problems in large-scale hybrid-parallel training that spans thousands of GPU servers and runs for weeks to months. Yet, these problems are not well studied, nor can they be quickly detected and effectively mitigated. In this paper, we first present a characterization study on a shared production cluster with over 10,000 GPUs1. We find that fail-slows are caused by various CPU/GPU computation and cross-node networking issues, lasting from tens of seconds to nearly ten hours, and collectively delaying the average job completion time by 1.34 current practice is to manually detect these fail-slows and simply treat them as fail-stops using a checkpoint-and-restart failover approach, which are labor-intensive and time-consuming. In this paper, we propose FALCON, a framework that rapidly identifies fail-slowed GPUs and/or communication links, and effectively tackles them with a novel multi-level mitigation mechanism, all without human intervention. We have applied FALCON to detect human-labeled fail-slows in a production cluster with over 99 further demonstrates that FALCON effectively handles manually injected fail-slows, mitigating the training slowdown by 60.1

Serverless computing platforms have become increasingly popular for running machine learning (ML) tasks due to their user-friendliness and decoupling from underlying infrastructure. However, auto-scaling to efficiently serve incoming requests still remains a challenge, especially for distributed ML training or inference jobs in a serverless GPU cluster. Distributed training and inference jobs are highly sensitive to resource configurations, and demand high model efficiency throughout their lifecycle. We propose FaPES, a FaaS-oriented Performance-aware Elastic Scaling system to enable efficient resource allocation in serverless platforms for ML jobs. FaPES enables flexible resource loaning between virtual clusters for running training and inference jobs. For running inference jobs, servers are reclaimed on demand with minimal preemption overhead to guarantee service level objective (SLO); for training jobs, optimal GPU allocation and model hyperparameters are jointly adapted based on an ML-based performance model and a resource usage prediction board, alleviating users from model tuning and resource specification. Evaluation on a 128-GPU testbed demonstrates up to 24.8% job completion time reduction and ×1.8 Goodput improvement, as compared to representative elastic scaling schemes.

Despite the advanced intelligence abilities of large language models (LLMs) in various applications, they still face significant computational and storage demands. Knowledge Distillation (KD) has emerged as an effective strategy to improve the performance of a smaller LLM (i.e., the student model) by transferring knowledge from a high-performing LLM (i.e., the teacher model). Prevailing techniques in LLM distillation typically use a black-box model API to generate high-quality pretrained and aligned datasets, or utilize white-box distillation by altering the loss function to better transfer knowledge from the teacher LLM. However, these methods ignore the knowledge differences between the student and teacher LLMs across domains. This results in excessive focus on domains with minimal performance gaps and insufficient attention to domains with large gaps, reducing overall performance. In this paper, we introduce a new LLM distillation framework called DDK, which dynamically adjusts the composition of the distillation dataset in a smooth manner according to the domain performance differences between the teacher and student models, making the distillation process more stable and effective. Extensive evaluations show that DDK significantly improves the performance of student models, outperforming both continuously pretrained baselines and existing knowledge distillation methods by a large margin.

Pipeline parallelism has been demonstrated to be a remarkable approach to improve throughput for training deep neural networks with billions of parameters over heterogeneous clusters. The 1F1B scheduling plan is a widely adopted strategy for memory and performance optimization, which interchanges the forward and backward stage computations of different micro-batches. On the other hand, a common issue in using the 1F1B scheduling is that stage computation is delayed due to the data transfer when network resources are preempted by other tasks, even with the minimum communication between stages. The exclusive access of these network resources cannot be guaranteed in cloud offerings. We present a general scheduling technique to accommodate pipeline parallelism to preempted network environments at the expense of a certain amount of memory pressure. The core concept is to extend 1F1B schedule scheme to kFkB, which groups k micro-batches, and alternately executes k forward and backward computations. We propose Ada-Grouper, an adaptive kFkB scheduler which regularly adjusts the number of group members k to maintain an optimal balance between communication and computation efficiency correspond to changes in a changing network environment under the memory limit. Experimental results demonstrate that our design maintain stable performance for pipeline parallelism, yielding a performance increase of up from 4% to 30%, compared with 1F1B in preempted network scenarios.

The development of personalized recommendation has significantly improved the accuracy of information matching and the revenue of e-commerce platforms. Recently, it has two trends: 1) recommender systems must be trained timely to cope with ever-growing new products and ever-changing user interests from online marketing and social network; 2) state-of-the-art recommendation models introduce deep neural network (DNN) modules to improve prediction accuracy. Traditional CPU-based recommender systems cannot meet these two trends, and GPU-centric training has become a trending approach. However, we observe that GPU devices in training recommender systems are underutilized, and they cannot attain an expected throughput improvement as what it has achieved in Computer Vision (CV) and Neural Language Processing (NLP) areas. This issue can be explained by two characteristics of these recommendation models: First, they contain up to a thousand of input feature fields, introducing fragmentary and memory-intensive operations; Second, the multiple constituent feature interaction submodules introduce substantial small-sized compute kernels. To remove this roadblock to the development of recommender systems, we propose a novel framework named PICASSO to accelerate the training of recommendation models on commodity hardware. Specifically, we conduct a systematic analysis to reveal the bottlenecks encountered in training recommendation models. We leverage the model structure and data distribution to unleash the potential of hardware through our packing, interleaving, and caching optimization. Experiments show that PICASSO increases the hardware utilization by an order of magnitude on the basis of state-of-the-art baselines and brings up to 6× throughput improvement for a variety of industrial recommendation models. Using the same hardware budget in production, PICASSO on average shortens the walltime of daily training tasks by 7 hours, significantly reducing the delay of continuous delivery.

Mixture-of-Experts (MoE) models can achieve promising results with outrageous large amount of parameters but constant computation cost, and thus it has become a trend in model scaling. Still it is a mystery how MoE layers bring quality gains by leveraging the parameters with sparse activation. In this work, we investigate several key factors in sparse expert models. We observe that load imbalance may not be a significant problem affecting model quality, contrary to the perspectives of recent studies, while the number of sparsely activated experts $k$ and expert capacity $C$ in top-$k$ routing can significantly make a difference in this context. Furthermore, we take a step forward to propose a simple method called expert prototyping that splits experts into different prototypes and applies $k$ top-$1$ routing. This strategy improves the model quality but maintains constant computational costs, and our further exploration on extremely large-scale models reflects that it is more effective in training larger models. We push the model scale to over $1$ trillion parameters and implement it on solely $480$ NVIDIA V100-32GB GPUs, in comparison with the recent SOTAs on $2048$ TPU cores. The proposed giant model achieves substantial speedup in convergence over the same-size baseline.

Pervasive needs for data explorations at all scales have populated modern distributed platforms with workloads of different characteristics. The growing complexities and diversities have thereafter imposed distinct challenges to execute them on shared clusters in corporate or public clouds. This paper presents Fangorn, an adaptive execution framework built on an enriched graph model. As the underlying infrastructure for core computation platforms at Alibaba, Fangorn supports various execution modes and caters to heterogeneous workloads. With the capability to orchestrate graph executions with both long-running and requested-on-demand resources at the same time, Fangorn allows exploration of tradeoffs between latency and resource efficiency, for jobs of all scales. By modeling distributed job executions as mutable graphs with pluggable components, Fangorn offers a systematic framework to adjust job executions adaptively, according to data statistics collected during run-time. Fangorn supports an array of different computation engines ranging from relational to deep learning, and is fully deployed on production clusters across Alibaba. It manages tens of millions of distributed jobs daily, with job size scaling from one to half-million. PVLDB Reference Format: Yingda Chen, Jiamang Wang, Yifeng Lu, Ying Han, Zhiqiang Lv, Xuebin Min, Hua Cai, Wei Zhang, Haochuan Fan, Chao Li, Tao Guan, Wei Lin, Yangqing Jia and Jingren Zhou. Fangorn: Adaptive Execution Framework for Heterogeneous Workloads on Shared Clusters. PVLDB, 14(12): 2972 -

Pervasive needs for data explorations at all scales have populated modern distributed platforms with workloads of different characteristics. The growing complexities and diversities have thereafter imposed distinct challenges to execute them on shared clusters in corporate or public clouds. This paper presents Fangorn, an adaptive execution framework built on an enriched graph model. As the underlying infrastructure for core computation platforms at Alibaba, Fangorn supports various execution modes and caters to heterogeneous workloads. With the capability to orchestrate graph executions with both long-running and requested-on-demand resources at the same time, Fangorn allows exploration of tradeoffs between latency and resource efficiency, for jobs of all scales. By modeling distributed job executions as mutable graphs with pluggable components, Fangorn offers a systematic framework to adjust job executions adaptively, according to data statistics collected during run-time. Fangorn supports an array of different computation engines ranging from relational to deep learning, and is fully deployed on production clusters across Alibaba. It manages tens of millions of distributed jobs daily, with job size scaling from one to half-million.

The pre-trained neural models have recently achieved impressive performance in understanding multimodal content. However, it is still very challenging to pre-train neural models for video and language understanding, especially for Chinese video-language data, due to the following reasons. Firstly, existing video-language pretraining algorithms mainly focus on the co-occurrence of words and video frames, but ignore other valuable semantic and structure information of video-language content, e.g., sequential order and spatiotemporal relationships. Secondly, there exist conflicts between video sentence alignment and other proxy tasks. Thirdly, there is a lack of large-scale and high-quality Chinese video-language datasets (e.g., including 10 million unique videos), which are the fundamental success conditions for pre-training techniques. In this work, we propose a novel video-language understanding framework named Victor, which stands for VIdeo-language understanding via Contrastive mulTimOdal pRe-training. Besides general proxy tasks such as masked language modeling, Victor constructs several novel proxy tasks under the contrastive learning paradigm, making the model be more robust and able to capture more complex multimodal semantic and structural relationships from different perspectives. Victor is trained on a large-scale Chinese video-language dataset, including over 10 million complete videos with corresponding high-quality textual descriptions. We apply the pre-trained Victor model to a series of downstream applications and demonstrate its superior performance, comparing against the state-of-the-art pre-training methods such as VideoBERT and UniVL.

Sparse expert models can achieve promising 001 results with outrageous large amount of param- 002 eters but constant computation cost, and thus 003 it has become a trend in model scaling. Still, 004 it is a mystery how Mixture-of-Experts (MoE) 005 layers leveraging the parameters with sparse 006 activation bring quality gains. In this work, we 007 investigate several key factors in sparse expert 008 models. We find that load imbalance may not 009 be a significant problem affecting model qual- 010 ity, and auxiliary balancing loss can be removed 011 without significant performance degrade. We 012 further discover that larger number of sparsely 013 activated experts k may not necessarily bene- 014 fit the performance on the time basis, and we 015 observe diminishing marginal utility that the 016 performance gap gradually narrows with the 017 increase in k We take a step forward to pro- 018 pose a simple method called expert prototyping 019 that splits experts into different prototypes and 020 applies top- k routing for each prototype in par- 021 allel. Our experiments demonstrate that the pro- 022 totyping strategy improves the model quality, in 023 comparison with further increasing to a larger k 024 with comparable computation cost to prototyp- 025 ing. Furthermore, we conduct an exploration on 026 training extremely large-scale models, and we 027 figure out that the strategy shows greater effec- 028 tiveness in training larger models. Notably, we 029 push the model scale to over 1 trillion parame- 030 ters on solely 480 NVIDIA V100-32GB GPUs. 031 The proposed giant model M6-T with expert 032 prototyping achieves substantial speedup in 033 convergence over the same-size baseline. 034

Apsara is a large-scale general-purpose distributed computing system developed at Alibaba Cloud Computing Inc (also called Aliyun). It is responsible for managing cluster resources within a datacenter, as well as scheduling parallel execution for a wide range of distributed online and offline applications. It is the common foundation of a majority of public cloud services offered by Aliyun, and supports all data processing workload within Alibaba as well. Some key components of the Apsara system include: (a) a distributed file system, called Pangu; (b) a distributed computation framework, called Fuxi, for cluster resource management and job scheduling; and (c) a parallel DAG-based data computation framework on top of Fuxi to process large datasets. Apsara is mostly written in C/C++ to achieve high performance and efficiency. More details about Apsara can be found in [1]. Apsara has been deployed on hundreds of thousands of physical servers across tens of data centers at Aliyun, with the largest clusters consisting of more than 5000 servers each. Any DAG-based distributed data processing can be implemented as a Fuxi job, including simple Map/Reduce jobs and more sophisticated machine learning jobs. Source inputs and outputs, along with intermediate data are placed in Pangu, with a variety of replication and locality configurations to achieve the optimal balance between performance and reliability in order to accommodate varying requirements from different applications. In this paper, we introduce FuxiSort, a distributed sort implementation on top of Apsara, and present our Daytona GraySort and MinuteSort results with details in technical implementation, performance improvements, and efficiency analysis. FuxiSort is able to complete the 100TB Daytona GraySort benchmark in 377 seconds on random non-skewed dataset and 510 seconds on skewed dataset, and Indy GraySort benchmark in 329 seconds. For the Daytona MinuteSort benchmark, FuxiSort sorts 7.7 TB in 58 seconds on random non-skewed dataset and 79 seconds on skewed dataset. For the Indy MinuteSort benchmark, FuxiSort sorts 11 TB in 59 seconds.